Vidu Q4 Preview debuts at #3 on Artificial Analysis AA-Video-I2V v1.0 leaderboard

All the information for this release comes from a single post by Artificial Analysis on its X account, which also included comparison examples from Video Arena. In other words, the two key categories of data—the ranking and the price comparison—are both provided by the leaderboard's publisher itself, belonging to its own evaluation system, and there is currently no other independent source for cross-verification.

The post did not disclose the model architecture, training data, or evaluation sample size, nor did it explain the evaluation methodology of the AA-Video-I2V v1.0 leaderboard or the total number of participating models, and it did not give the specific score gap between Vidu Q4 Preview and the No. 1 and No. 2 entries. These gaps mean all the numbers below should be understood as "results under the publisher's own account," not conclusions that have been reviewed by multiple parties.

According to the post, Vidu Q4 Preview debuted at No. 3 on the AA-Video-I2V v1.0 image-to-video leaderboard; on the same leaderboard, the previous-generation Vidu Q3 Pro ranks No. 19. The two are 16 places apart, and Vidu Q4 Preview thus enters the top three of that leaderboard.

This comparison is treated as a core signal because it places two generations of models under the same evaluation system for direct ranking. One commenter said outright that going from No. 19 to No. 3 is the only important information, and the rest is just better packaging of latency and bitrate—this is an outside commenter's view, not a verifiable conclusion.

The questions that need to be retained are equally clear: who the No. 1 and No. 2 entries are, and how large the gap is between Vidu Q4 Preview and them, are not stated in the post; nor is it clear whether the result can be reproduced by other third-party evaluations.

On the capability side, the parameters given in the post are relatively specific: Vidu Q4 Preview can generate 3- to 16-second clips, with a maximum resolution of 4K, and generates video with native audio.

In Reference to Video mode, it accepts up to 15 image references and 3 audio references. These numbers describe the upper limits on the input side, not output quality itself, so they are better understood as usage constraints rather than evidence of image quality.

Price is the most easily verifiable part of this release: Vidu Q4 Preview is priced the same as Vidu Q3 Pro, at 24 credits per second for 1080p video with audio, about $7.20 per minute at Vidu's standard rates.

The accompanying conclusion is that Artificial Analysis says it offers better image quality at the same price. This statement is the publisher's claim; the post provides no specific metrics or scores to support it, so it cannot be quantified; nor does the post explain the conversion between 24 credits/second and $7.20/minute or whether the rates will remain valid long term.

So the more prudent reading is: the price being unchanged is a confirmable fact, while "better image quality" is currently only the publisher's judgment.

Vidu Q4 Preview is Vidu's new video model, released ahead of the full Q4 version, and is currently available via the Vidu API and the Vidu website.

The preview identity brings two unresolved questions: whether the ranking will change after the full Q4 version is released, and whether the preview and final versions will be consistent in capability and price—the post gives no answers to either.

To gather the verifiable parts: Vidu Q4 Preview debuted at No. 3, Vidu Q3 Pro ranks No. 19, a gap of 16 places; the model supports 3 to 16 seconds, up to 4K, native audio, and in Reference to Video up to 15 image references and 3 audio references; pricing is unchanged from the previous generation, at 24 credits per second for 1080p video with audio, about $7.20 per minute; available via the Vidu API and website.

The parts still to be completed are equally clear: the image-quality improvement lacks quantitative basis, the leaderboard methodology and scale of participation are undisclosed, the gap to the top is unknown, the reproducibility of the results is unverified, and the relationship between the preview and final versions is unexplained. Until these questions are answered, No. 3 is a ranking worth recording, not a completed conclusion.